How a network computes, why gradients vanish, and what makes depth trainable
The regularizers you already have shrink weights. optim.11 adds a penalty on their size, stats.18 does the same for regression, and both work by making large weights expensive. Dropout does something that sounds reckless instead: during training it switches hidden units off at random and lets the rest carry on without them.
🔒 This is a Pro lesson — the interactive figure, worked examples, quiz and practice open with Pro access.
▶ Dropout